Non-invasive Diabetes Detection using Facial Texture Features Captured in a Less Restrictive Environment

• Main Authors: Garcia, Christina A, Reyes, Rosula SJ, Abu, Patricia Angela R
• Format: text
• Published: Archīum Ateneo 2019
• Subjects: Diabetes Mellitus; Gabor Filter; Texture Features.; Biomedical; Electrical and Computer Engineering
• Online Access: https://archium.ateneo.edu/ecce-faculty-pubs/39; https://www.ijrte.org/wp-content/uploads/papers/v8i2S11/B15840982S1119.pdf
• Institution: Ateneo De Manila University

The prevalence of Diabetes Mellitus (DM) worldwide has risen dramatically with 1 of 3 deaths happening in Western Pacific region according to the 2017 report of International Diabetes Federation. The Philippines ranks 5th in WP with the most cases of diabetes. Local experts and IDF estimate that half of the people with diabetes are unaware they have it and will likely remain undiagnosed. Conventional ways to detect if a person has diabetes are often invasive and painful such as puncturing fingers for blood sample. Though non-invasive DM detection techniques have gained consideration in more analysts, presently they have restrictive set-up for image capture. This paper explores the performance of using mobile device as a convenient tool for image capture of DM and healthy dataset for non-invasive detection using facial block texture features and Gabor filter. Filipino participants that undergo regular check-ups for diabetes monitoring were chosen within the age inclusion criteria of 20 to 79 years old in which surveys for Philippines assessed the occurrence of diabetes to be most prevalent according to IDF and World Health reports. For each subject, a mobile device 12mp and 7mp cameras were used to take the photo placed 30 cm in front of the face under normal lighting condition to ensure full coverage and avoid unnecessary background. A ratio of 70:30 training to testing set was maintained and extracted facial blocks were classified using SVM and KNN. A total of 100 images from each camera were captured, preprocessed, filtered and iterated to compare performance of data. 90% accuracy, 93% sensitivity and 93% specificity were achieved for 12mp with SVM. For the 7mp camera, an accuracy of 80% using SVM and 93% sensitivity using KNN were achieved after increasing the predictors obtained for classification.